Under the soak workload V8's compiled code and the layers' caches grow for about 250 switches and then the heap is flat (121.4, 121.7, 120.4, 119.1 MiB after 250 to 400 switches), so a line fitted from switch 20 measured the warm-up. Over the second half the threshold stays 10 KiB per switch; the run before the layer effect fix measures 19 KiB there.
Every scenario runs through scripts/qa-strict-harness.mjs: each page, frame and worker gets an ordered block list before it runs (paid hosts and routes first, then only the preview, declared fixtures, keyless probes and cached free tile hosts), switches with real input on the picker, and records its assertions per layer and engine. The runner reads the catalog from the live page, fails any layer without an assertion on either engine, can build and serve a keyless checkout itself, and adds rapid, hidden-tab, context-loss, failure-injection, cold-chunk and soak scenarios. The main chunk gets a 1% size budget.